When multiple signals are present, it is often desirable to separate the signals. For example, a microphone may pick up speech signals from multiple individuals, thereby generating a mixed signal. However, one skilled in the art can imagine scenarios where it may be desirable to extract a single user's speech from the mixed signal.
Previous approaches to extracting signals from mixtures without knowledge of how the mixtures were formed, otherwise known as blind source separation (BSS), have all been based on measuring and adjusting the low-level properties of the output signals themselves (such as the statistics of the signal waveforms). The prior art approach for blindly separating signals from mixtures gathered using multiple sensors is shown in FIG. 1. It is a feed-forward architecture in which various BSS methods, such as independent component analysis (ICA) or the constant modulus algorithm, are used to extract signals from mixtures using the low-level statistical properties of the signals. These statistical properties can include statistical independence, kurtosis, sparsity, constant modulus, and other signal-level properties. For example, ICA extracts signals from mixtures by maximizing the statistical independence of the outputs. Alternatively, a constant modulus algorithm assumes the source signals have a constant modulus and extracts signals by forcing the outputs to have a constant modulus.
Relying on such low-level properties of the signals for signal-extraction has several disadvantages. Long signal lengths are required in order to accurately estimate the signal properties. Additionally, many conventional BSS algorithms are based on gradient descent, which relies on good continuous estimates of the signal properties. Therefore, BSS methods have difficulty in following fast changes in the mixing channels. Many conventional BSS algorithms assume certain properties of the signals which are not found in practice or severely limit the range of signals that can be handled.
BSS based on signal statistics has been successfully applied in a number of applications, such as blind antenna beamforming and removal of interference from speech signals. However, the quality of the signal separation is limited by the accuracy of the estimates of the signal statistics. In many cases, high accuracy requires the collection and processing of large amounts of signal data. This results in slow update rates that may not follow rapid changes in the mixtures due both to the processing requirements and the need for long signal lengths for good statistics.
BSS was further described by F. Rojas, I. Rojas, R. M. Clemente, and C. G. Puntoner, in a paper entitled, “Nonlinear Blind Source Separation Using Genetic Algorithms.”The paper was published in Proceedings of International Conference on Independent Component Analysis, 2001. In the aforementioned paper, the authors describe using genetic algorithms to separate signals. However, their method is based on the low-level statistical properties of the signals (statistical moments), not on the cognitive information encoded in the signals.
While cognitive information exists in the signals, no known prior art extracts the signals based on the cognitive information encoded in the signals. In cognitive signal separation, signals can be extracted based on the performance of a higher-level system that actually recognizes and uses the cognitive information in the signal. Such a system would greatly increase the range of signals that can be processed and improve system performance because the signal-extraction is tied to the end performance of the system, not to artificially-selected low-level properties of the signals that may be irrelevant to the top-level end-user of the cognitive information contained in the signals. For example, ICA will fail if the source signal statistics are too close to Gaussian because the higher-order moments will be too small and noisy. Cognitive signal separation is blind to the low-level signal statistics so it can handle such a case. As opposed to BSS systems, a cognitive-based system does not require gradient information or continuity in the classifier confidence as a function of extraction parameters. A cognitive-based system can also be adapted to handle nonlinear and convolutive mixtures of signals.
As described above, the prior art extracts signals by measuring and adjusting the low-level properties of the output signals themselves as opposed to using the higher-level information encoded in the source signals. Thus, a continuing need exists for a cognitive-based system that allows a user to process signals using higher-level information encoded in the source signals themselves.